Apache Kafka

Quality: Silver Tier.

🐙 userver provides access to Apache Kafka Brokers via two independent interfaces: kafka::ProducerComponent and kafka::ConsumerComponent.

They expose APIs for sending messages to Kafka topics, suspending coroutine until message delivery succeeded or failed, and for process message batches from the subscribed set of topics, asynchronously polling messages in separate task processor.

Common Features

• Easy to configure (compared with raw Kafka clients with dozens of options);
• No message payload copying;
• Support of SASL SCRAM-SHA-512 authentication and SSL transport for Broker communication;
• Comprehensive logs of all events and errors;
• Metrics;
• Kafka message headers support;

Producer Features

• 🚀 Parallel cooperative messages delivery reports processing (in comparison with all librdkafka-based Kafka clients, the performance of which rests on a single thread);
• 🚀 No blocking waits in implementation (message senders suspend their coroutines execution until delivery reports occurred);
• Synchronous and asynchronous non-blocking interfaces for producing messages;
• Automatic retries of transient errors;
• Support of idempotent producer (exactly-once semantics);
• Sending message to concrete topic's partition;

Consumer Features

• 🚀 No blocking waits in implementation (message poller suspends the coroutine until new events occurred);
• Callback interface for handling message batches polled from subscribed topics;
• Balanced consumer groups support;
• Automatic rollback to last committed message when batch processing failed;
• Partition offsets asynchronous commit;

Metrics

Producer and consumer yields same set of metrics (but with different prefixes):

Metric name	Labels	Description
kafka_producer.avg_ms_spent_time	component-name, topic	Average message producer latency
kafka_producer.messages_total	component-name, topic	Total number of sent messages
kafka_producer.messages_success	component-name, topic	Number of messages successfully written to partition
kafka_producer.messages_error	component-name, topic	Number of messages failed to produce
kafka_producer.connections_error	component-name	Number of broker connection errors occured
kafka_consumer.avg_ms_spent_time	component-name, topic	Average time between message written to partition and read by consumer
kafka_consumer.messages_total	component-name, topic	Total number of read messages
kafka_consumer.messages_success	component-name, topic	Number of successfully processed messages
kafka_consumer.messages_error	component-name, topic	Number of messages user-callback failed
kafka_consumer.connections_error	component-name	Number of broker connection errors occured

See Service Statistics and Metrics (Prometheus/Graphite/...) for info on how to get the metrics.

Planned Enhancements

• ✅ Transfer from raw polling with timeouts to events processing, making the message polling non-blocking and leading to better library scalability;
• ✅ testsuite Kafka support in OSS;
• Support of different compression codecs (GZIP, LZ4, ZSTD, etc..);
• Support more SASL authentication mechanisms (GSSAPI, OAUTHBEARER);

Unit and functional testing

• Kafka service shows how to test userver services with Kafka consumers and producers;
• kafka::utest::KafkaCluster may be used for convenient unit testing with local consumers and producers instances;

More information

• For configuration see kafka::ProducerComponent and kafka::ConsumerComponent;
• For understanding of interfaces see kafka::Producer and kafka::ConsumerScope;
• Also kafka::impl::Consumer and kafka::impl::Configuration are well documented and may be useful;

---

⇦ MySQL driver design and implementation details | YDB ⇨